Haswell vs Kaveri

[homerdog]

Only 4 nucleos? I don't really follow everything you guys have been saying here, but I assume these nucleos are way better than the ones in my 3550?

 

[Alexko]

I think the consensus is that they're expected to be about 10% better.

 

[iMacmatician]

From an AMA by jecb on Reddit, "IAmA CPU Architect and Designer at Intel, AMA" via XtremeSystems forums and 3DCenter forums.

I've been involved in many of Intel's flagship processors from the past few years and working on the next generation. More specifically, Nehalem (45nm), Westmere (32nm), Haswell (22nm), and Broadwell (14nm).

In technical aspects, I've been involved in planning, architecture, logic design, circuit design, layout, pre- and post-silicon validation. I've also been involved in hiring and liaising with university research groups.

This person answered many questions, and in doing so, gave some hints on Haswell and Broadwell,

If you like to overclock, Haswell is worth it (can't tell you why but read the Haswell Anandtech preview very carefully for buried treasure). On-die graphics is improving quite a bit as well. If you're into energy efficiency or even more graphics, Broadwell. I think the tech community will be very pleasantly surprised with Broadwell.

comments on the rumor about Haswell CPUs being soldered,

This rumor is likely misinterpreting facts or based on really incomplete information. Many of the variants will be BGA packages for certain form factors but not all.

and

it's clear that the people who leak to semiaccurate work very far removed from any real engineering work. We have a laugh when they get code names wrong or announce features that don't exist.

 

[Homeles]

The only mention of anything overclocking related, that I saw in the AnandTech article, was a comment about an improved memory controller and Intel nudging DRAM manufaturers to push for higher frequencies.

In comparison to the Ivy Bridge architecture detailing Anand did, it seems that Haswell will bring more substantial memory controller improvements. In the Ivy Bridge, he wrote that the IMC is "relatively unchanged." With the Haswell article, he very, very briefly mentions that the memory controller is "improved, with better write throughput."

Ivy Bridge already supports ridiculous DDR3 speeds; perhaps Haswell will bring us to ludicrous speed?

I think the consensus is that they're expected to be about 10% better.

In existing applications, yes. There's a lot of changes that seem they would make a large difference in existing applications -- better branch prediction, the execution engine widening, etc. My theory is that the L3 latency and clock regression is what is holding back larger performance numbers.

The most interesting changes that Haswell makes are in support of future applications though, so here's to hoping Haswell manages to see significant consumer -- and therefore developer -- adoption.

 

[UniversalTruth]

Top Richland 28nm APU is A10 6800K

The expected launch date is June, but the production should be ready by late March 2013. Production candidate samples were scheduled for early December, while late January should be the time for production ready samples.

With some interesting comments following the article.
Accurate or semi-accurate?

 

[DavidC]

In existing applications, yes. There's a lot of changes that seem they would make a large difference in existing applications -- better branch prediction, the execution engine widening, etc. My theory is that the L3 latency and clock regression is what is holding back larger performance numbers.

Actually according to rumors it was Intel was expecting 10% at the minimum. The 10% number is actually similar to what was said with Sandy Bridge.

 

[Homeles]

Actually according to rumors it was Intel was expecting 10% at the minimum. The 10% number is actually similar to what was said with Sandy Bridge.

As much as I would like to believe this, I can't find any sources that verify your Sandy Bridge performance claim.

Also, I haven't seen any Intel-sourced estimates on performance improvement percentages for Haswell. The numbers I've seen tossed around were 5-15% at AnandTech, or "10-20% at best" over at XbitLabs, or 10% on average as estimated by Real World Technologies. Perhaps they got the numbers from Intel, but I haven't seen Intel directly cited, and it seems that each site has claimed their performance numbers as their own conjecture, rather than being influenced by what Intel has claimed.

 

[iMacmatician]

Thanks to Olivon (XS forums)

From PConline: "Dominate: Intel 4th generation Core i series comprehensive Secret" (original).

"Haswell" characteristics can be summarized as the following four points: 1,22 nm process new architecture, higher performance, greater overclocking potential, and integrated voltage regulator; 2, a new instruction set, Haswell Add new AVX instruction set, to improve the performance of AES-NI; 3, Core Graphics enhanced support DX11.1 OpenCL1.2 optimize 3D performance, support HDMI, DP, DVI, VGA interface standard; 4, the interface changes, use LGA1150 interface, is not compatible with the old platform.

Official PDF documents, Intel refers to the fourth generation Core i series has more headroom for overclocking with more flexible adjustment, which is perhaps the fourth-generation Core i integrate the results of the complete voltage regulator.

From PC Watch: "'Integrated Voltage Regulator' strongest weapon 'Haswell'" (original).

There's also a neighboring article "Way of integration of the voltage regulator to Haswell from "Ozette" chip of Intel" (original).

"Ozette" appears to be some sort of prototype chip (from 2007) involving voltage control.

 

[UniversalTruth]

I think the consensus is that they're expected to be about 10% better.

10% faster is a joke, isn't it? I mean it is so negligible that it is not worth it to even discuss it... not to mention considering upgrades or something more serious.

 

[rpg.314]

10% faster is a joke, isn't it? I mean it is so negligible that it is not worth it to even discuss it... not to mention considering upgrades or something more serious.

5-10%/yr in serial performance is all you are going to get. I thought that was common knowlege since Netburst.

 

[Homeles]

10% faster is a joke, isn't it? I mean it is so negligible that it is not worth it to even discuss it... not to mention considering upgrades or something more serious.

In environments where performance actually matters, the +10% benefit in existing applications is a non-issue, because the real benefits of Haswell lie elsewhere. If Intel (or any other computer architecture firm) only cared about legacy performance, the computing world would be a sad place. Thankfully, that's not how things work, and we have things like AVX2 and TSX to look forward to.

 

[DuckThor Evil]

Sounds like a price hike for the top i5 model, since it goes from x570 to x670.

 

[lanek]

5-10%/yr in serial performance is all you are going to get. I thought that was common knowlege since Netburst.

If only AMD could push 15% 2 times in follow for coming back...

 

[Homeles]

If only AMD could push 15% 2 times in follow for coming back...

Steamroller should bring that alone... if it ever releases.

 

[UniversalTruth]

Steamroller should bring that alone... if it ever releases.

Or? What is the other option? AMD should release something after all, even if it is not called Steamroller. Or else...
I think AMD have their chance and it will be here with this low care from Intel for this particular part of the market.

Ok, the problem for Intel is that the market isn't growing and they want to grow and that's why they push their attention in other directions.
On the other hand, AMD has the potential to occupy what Intel abandons, so, I think it would be a huge mistake from AMD to follow other markets with higher priority.

 

[liolio]

5-10%/yr in serial performance is all you are going to get. I thought that was common knowlege since Netburst.

Do you think the jump could be bigger than the increase in serial performance?

I wonder about the impact of the choices made by Intel on multi threaded scenario, like games.
I mean in a lot of games the Core i3 were doing great but they were case we they lose to AMD quad cores for example.

Now Haswell can issue instruction to 2 SIMD/FP units along with other improvements, I for one expect pretty big improvements at legacy applications/ games.
I think that for gamers at least, the next core i3 should provide a beefy jump in perfs, especially the heavily multi threaded games as the 2 threads should have more resources to breath.
Actually looking at the rumors in the consoles realm and how consoles specs affect games development, I would not be too surprised if the up coming core i3 (haswell and Broadwell) could be all a gamer needs as far as gaming is concerned (super enthusiast aside).

So looking at the kind of review people are interested in here (so review that involved games testing) I think that Haswell, especially the core i3 line, should bring beefy improvements vs previous products.

Then that is for "legacy" games, Haswell comes with other goodies that may make quiet a difference if futures games take advantage of it.

 

[rpg.314]

Do you think the jump could be bigger than the increase in serial performance?

AVX2 will not used pervasively, even by games. So it is not going to be much better than that.

 

[liolio]

AVX2 will not used pervasively, even by games. So it is not going to be much better than that.

I'm not speaking of avx2 except in my last sentences. I speak about the fact that on existing code SSE/AVx the new core can dispatch 2 SIMD instruction per cycle on top of others improvements which got me to think that on "legacy" games (so existing games and obviously not using avx 2 and quiet possibly not using avx) along with the increases/improvements made elsewhere I expect greater than 10% improvements. I think especially games where the 4 hardware thread of a core i3 could not trump for example the four "real" cores of say a phenom X4.

 

[3dilettante]

Could you clarify what you mean?
Prior cores could issue two SIMD ALU instructions.

There is one area where Haswell may improve in legacy code, if the instruction mix is SIMD MUL heavy.
The SIMD ADD capacity did not double.

edit: Or that might be a 256-bit mode restriction, I'm not certain.

 

[liolio]

Could you clarify what you mean?
Prior cores could issue two SIMD ALU instructions.

There is one area where Haswell may improve in legacy code, if the instruction mix is SIMD MUL heavy.
The SIMD ADD capacity did not double.

I meant no more than what you said
It is a bit over my head but it is about what D. Kanter discusses here:
http://www.realworldtech.com/haswell-cpu/4/
So more executions ports but also more 'resources', and how that going to affect existing multi-threaded code. As to me it looks like the 2 thread will have more room to play / the core will have more room to dispatch instruction of the two cores.
So you kind of answered me wrt to the SIMD, as they added only another Multiply alu (FMA doesn't exist before AVX).
Though I still wonder about the overall "results" on multi threaded code and if could get greater than the 10% increase in single thread performance, I think it should though I do not the extend.

EDIT
D. Kanter says, and he knows what he is speaking about, that existing applications should see a boost of ~10%. But looking at the scope of applications he might be looking at I wonder what would be the impact on for example games /multimedia apps. I would think that when it says 10% that includes plenty of workloads that may not or barely touch the SIMD.